Generative Artificial Intelligence models that design new materials typically sample structures from large training datasets and tend to favor stable, common outcomes. That bias limits discovery of materials with exotic quantum properties, such as candidates for quantum spin liquids or platforms for high-temperature superconductivity. To address that gap, researchers at MIT created SCIGEN, short for Structural Constraint Integration in GENerative model, a code that enforces user-defined geometric constraints at each iterative step of diffusion-model generation so the model produces structures with specific lattice patterns.
SCIGEN was evaluated by integrating it with a popular materials generation model known as DiffCSP and directing the model to create structures with Archimedean lattices, including Kagome and Lieb patterns linked to quantum phenomena and flat bands. The constrained model produced more than 10 million candidate materials; one million passed an initial stability screen. The team used Oak Ridge National Laboratory supercomputers to run detailed atomistic simulations on a 26,000-sample subset and found magnetic behavior in 41 percent of those structures. From the most promising outputs, researchers synthesized two previously unknown compounds, TiPdBi and TiPbSb, in experimental labs run by Weiwei Xie and Robert Cava, and subsequent measurements largely agreed with the model predictions.
The work, described in a paper published in Nature Materials, is led by Mingda Li with co-authors from MIT, Emory University, Michigan State University, Oak Ridge National Laboratory, and Princeton University. The researchers highlight that SCIGEN shifts generative models away from prioritizing only stability and toward design rules valued by quantum materials researchers, supplying experimentalists with many more candidates to test. They note that laboratory synthesis and characterization remain essential and that future directions include adding chemical and functional constraints to generative models. The project received support from the U.S. Department of Energy, the National Energy Research Scientific Computing Center, the National Science Foundation, and Oak Ridge National Laboratory.
